Gemini 4 Argon ships to cyber defenders without guardrails

Gemini 4 Argon is not something you can sign up for. Google is releasing its new frontier model first to vetted cyber defenders in the Fairwind Program, and for that group plus its own internal teams it arrives with the cyber guardrails removed. Wider access comes later, starting with paid API customers and Google AI Ultra subscribers. The announcement is dated September 30, 2026.

Who gets it, and in what order

Google describes the rollout as staged: Fairwind participants first, then a gradual expansion coordinated with the U.S. government, then developers, enterprises and consumers. No date is attached to any of those steps. That makes the headline capability and the availability story two separate things.

Standard versions of the model keep their misuse safeguards. The unguarded build exists so that the model can hunt and patch vulnerabilities on its own, which is exactly the capability Google is trying to keep inside a vetted circle.

The dual-use problem

A model that finds flaws without being told where to look is as useful to an attacker as to a defender. Google lists four mitigations: refusing harmful cyber and CBRN requests while preserving dual-use research, what it calls leading robustness against prompt injection, misalignment monitoring over chains of thought, and hardened sandbox environments. Each of those rests on Google's own testing.

The timing is awkward. heise online points back to a configuration error in which Gemini reached real company systems through internet access that was left open by accident. Stripping the cyber restrictions raises the cost of that class of mistake.

Price and limits

Introductory rates are $2 per million input tokens and $10 per million output tokens, moving to $4 and $20 once that phase ends. Cached input tokens are discounted 95% against the input price.

The headline technical change is the output ceiling, which goes from 64,000 tokens to 1 million. The context window also sits at 1 million tokens.

Where it wins, and where Google stays quiet

Google claims a tie for first on CWE-bench v1 at 68% and the top spot on AutomationBench at 51.3%. Every figure in the table is vendor-reported.

heise online reads the selection as curated: the tests Google foregrounds are the ones Argon leads. On Terminal-Bench 4.0 it scores 57.4%, trailing Claude Sonnet 5.5 at 70.6% at broadly comparable pricing.

Three in-house claims nobody can audit

Google says it already runs Argon internally for quantum algorithm optimization, for memory optimization across its data centers, and for porting C/C++ code to Rust. The figures given are more than 300 TiB of memory freed and migrations covering more than 800,000 lines of code.

None of the three can be checked from outside. Treat them as vendor claims until independent measurements exist.